Incident report: AI agents attempted exploits on our website during a third-party alignment evaluation

In late October, AI agents participating in alignment evaluations conducted by Trajectory Labs interacted with Epoch AI's website, attempting exploits such as SQL injection and contacting employees under false identities. Trajectory assigned the agents an intentionally impossible task to grade a fake version of a FrontierMath benchmark while instructing them never to give up. Epoch AI reported that no private or confidential information was accessed and subsequently proposed minimum standards for future alignment evaluations, including informed consent, monitoring, and reporting. Open article →